a statistical method for the prediction of the loop tack

HAL Id: hal-00176886https://hal.archives-ouvertes.fr/hal-00176886

Submitted on 19 Mar 2013

HAL is a multi-disciplinary open accessarchive for the deposit and dissemination of sci-entific research documents, whether they are pub-lished or not. The documents may come fromteaching and research institutions in France orabroad, or from public or private research centers.

L’archive ouverte pluridisciplinaire HAL, estdestinée au dépôt et à la diffusion de documentsscientifiques de niveau recherche, publiés ou non,émanant des établissements d’enseignement et derecherche français ou étrangers, des laboratoirespublics ou privés.

A Statistical Method for the Prediction of the Loop Tackand the Peel of PSAs from Probe test Measurements

Isabelle Rivals, Léon Personnaz, Costantino Creton, François Simal, PatriceRoose, Steven van Es

To cite this version:Isabelle Rivals, Léon Personnaz, Costantino Creton, François Simal, Patrice Roose, et al.. A StatisticalMethod for the Prediction of the Loop Tack and the Peel of PSAs from Probe test Measurements.Measurement Science and Technology, IOP Publishing, 2005, 16, pp.2020-2029. �10.1088/0957-0233-16-10-018�. �hal-00176886�

https://hal.archives-ouvertes.fr/hal-00176886

https://hal.archives-ouvertes.fr

A STATISTICAL METHOD FOR THE PREDICTION OF THE LOOP TACK

AND THE PEEL OF PSAS FROM PROBE TEST MEASUREMENTS

Isabelle Rivals1, Léon Personnaz1, Costantino Creton2, François Simal3,

Patrice Roose3 and Steven van Es3

1Équipe de Statistique Appliquée, ESPCI, Paris, France.2Laboratoire de Physico-Chimie des Polymères et des Milieux Dispersés, ESPCI, Paris,

France.3R&D Surface Specialties, Drogenbos, Belgium.

`

Abstract

We investigated the potential of the probe test as a high throughput test for the rapid screening

of a large number of candidate pressure-sensitive-adhesives. The output of a successful

screening tool should be usable to predict with some precision important characteristics of

adhesives such as loop tack and peel which take longer to measure. The output of an

instrumented probe tack test being both reproducible and sensitive to changes in polymer

structure or formulation, we developed a statistical procedure which builds a polynomial model

of the force values at the relevant time instants only, the number of necessary monomials being

established using a test for lack of fit. The performance of the models is then estimated using

cross-validation and an independent test set. The prediction results obtained on a data set

representative of commercial adhesives show that the force signal recorded during a probe test

indeed contains exploitable information about the loop tack and peel forces, and that the

proposed statistical procedure is more successful for quantitative predictions than existing

alternative approaches.

Empirical prediction of loop tack and peel from probe tests measurements2

I. Introduction

Pressure-sensitive-adhesives (PSAs) are increasingly popular for fastening applications due to

their safe and easy handling. With the increase in the number and variety of applications come

more demanding requirements in terms of properties. In order to speed up the optimization of

the properties needed for a specific application, it will be become more and more important to

develop a good high throughput test for the rapid screening of a large number of materials, so

as to select only the most promising ones [Grunlan et al. 2004] [Kulikov et al. 2005] [Zhang et

al. 2005].

Initially, the instrumented probe tack test has been developed to gain more physical insight in

the mechanisms of debonding of PSAs [Zosel 1985] [Lakrout et al. 1999] [Shull & Creton

2004]. A careful analysis of the results obtained with the probe tack test has led to significant

advances in the understanding of the micromechanisms of debonding of a PSA layer from a

rigid substrate, and the main results can be found in a comprehensive review [Shull & Creton

2004]. Most of these advances came from the very detailed information obtained from the probe

test, which provides an entire stress-strain curve and not a unique value, coupled with an in situ

observation of the deformation mechanisms of the adhesive layer with a video camera [Creton et

al. 2001] [Brown et al. 2002] [Creton 2003].

These results have shown that for a very confined PSA layer which is being deformed in

traction, failure is initiated by the formation of cavities at the interface with the probe. These

cavities then form a foamed structure, which can be easily deformed to large strains in the

tensile direction. Final detachment of the adhesive from the probe occurs when the nonlinear

elastic properties of the adhesive display a significant strain hardening due to the finite

extensibility of the polymer chains forming the backbone of the adhesive.

Despite its obvious advantages, such as speed of execution and reproducibility [Chuang et al.

1997], the probe test is not yet widely accepted in industry, where adhesive properties are still

typically tested with standardized industry tests closer to applications, such as loop tack, peel or

shear tests [PSTC 2000]. It would therefore be economically advantageous to be able to predict

the result of a loop tack, peel or shear test, from a simple probe test experiment which typically

lasts less than a minute. However, the important conceptual advances that the more fundamental

investigations brought, did not yet result in a quantitative correlation between the outcome of the

probe test (i.e. a curve of stress as a function of strain) and a value such as loop tack, peel force

or shear, and further work is needed in that direction.

In this paper, we explore a different but parallel approach to predict the property of interest,

loop tack or peel value, from a probe test curve. This approach is based on the assumption that

the information needed to predict the property is contained in the curve obtained with the probe

tack test, but needs to be extracted with statistical tools. From the experimental point of view,

the requirements are simple: the probe test must be sensitive to changes in polymer structure and

Empirical prediction of loop tack and peel from probe tests measurements 3

formulation that will have an effect on the application property, and the probe test must be

reproducible to provide reliable data. The sensitivity of the test to changes in chemical structure

and in formulation has been shown in several publications on the subject [Zosel 1985] [Brown

et al. 2002] [Lakrout et al. 2001], and is summarized in Figure 1 which shows a typical probe

test curve for a PSA. The compression and initial part of the tensile curve are sensitive to the

linear viscoelastic properties of the adhesive; the peak in tension and the plateau following the

peak are much more sensitive to both adhesive properties and non-linear elastic properties.

Unlike a simple rheology experiment in the linear viscoelastic regime, the probe test deforms the

adhesive layer, in both the linear (compression) and the non-linear (tension) regime which will

be essential to predict debonding energies [Roos & Creton 2004]. Reproducibility is usually

better than for standard PSA tests, but can be further optimized by a careful sample preparation

and by the choice of a hemispherical probe, which is insensitive to small misalignments

[Chuang et al. 1997] [Crosby & Shull 1997].

Therefore a suitable statistical method should be able to extract the most relevant information

from the probe test curves and use them to build a predictive model of a specific property such

as loop tack or peel value. It is this path that we explore in our paper. The methodology we use

has been developed to extract the most significant part of a data set to build a predictive model

of the desired property, see [Rivals & Personnaz 2003a,b].

II. Experimental

Probe tests were performed for 26 different materials representative of commercial permanent

PSAs based on SIS (Kraton D-1160 and D-1161, from Kraton Polymers) and tackifing resins

(Piccotac 95E, Piccotac 212 and Foral 85-E, from Eastman Chemical Company), see Appendix

1. Three to five repeat tests were made for each material in the same conditions. The probe tests

were performed on a TA-XT2 i HR texture analyzer (Stable Microsystems) fitted with a

spherical probe having a diameter of 25 mm. The probe was brought in contact with the

adhesive layer at a velocity of 10 µm/s, until a compressive force of 1.1 N was reached. It was

kept in contact for 0.5 second at this load, and next removed at a debonding speed of 10 µm/s.

The whole force-distance curve (compression and traction) was recorded.

The adhesive films were prepared by transfer coating: a toluene solution of polymer and

tackifier was applied on Silicone paper using an automatic bar coater. The drying conditions

were as follows: 30 minutes at room temperature, followed by 3 minutes at 110°C. The dried

PSA film (thickness 30 mm) was then coated with a PET film (Mylar, 23 µm thick). The

laminate was conditioned for 24 hours under controlled humidity and temperature (23°C ± 2°C,

50% RH ± 5% RH) before testing.


0 10 20 30 40 50 60

-1

-0.5

0

0.5

1

1.5

time (s)

forc

e (

N)

Figure 1. Typical force curves obtained with four probe test repeats (curves N°300). The loop tack test

provides a mean value of 21.2 N/25 mm, and the peel test a mean value of 15.5 N/25 mm.

The peel tests were performed using FINAT FTM 1 conditions: 25 mm x 90 mm strips of PSA

were applied on stainless steel plates using a 1 kg FINAT standard test roller, and 180° peel

were measured at a peel rate of 300 mm/min, after 20 minutes. The loop tack tests were

performed using FINAT FTM 9 conditions: 25 mm x 210 mm strips of PSA were applied on

stainless steel plates. The loop was brought in contact and immediately removed at 300

mm/min. No stick-slip phenomenon was observed during the peel and loop tack tests.

The outcome of the experimental part of the study is a series of force curves with the

corresponding values of the peel force and of the loop tack. Typical force curves obtained with

4 probe test repeats are shown on Figure 1, with the mean values of the loop tack and of the

peel. Note that each repeat was performed on a fresh sample. The whole data set is described in

Appendix 2.

III. Design of the predictive models

This section is devoted to the design of predictive models of the loop tack and of the peel from

force descriptors, using the previous data set. Briefly, there are five steps: 1) the study of the

variability of the loop tack and peel measurements ; 2) the choice of the candidate descriptors,

i.e. of the variables that potentially contain the information about the loop tack and the peel ;

3) the construction of nested models of increasing complexity involving the candidate

descriptors and/ or fixed functions of them ; 4) the selection of the model of minimal necessary

complexity given the variability of the loop tack and peel values, i.e. involving only the most

relevant descriptors ; 5) the estimation of the performance of the model using both cross-

validation and an independent test set.

III.1. Variability of the loop tack and of the peel measurements

The data set involves 26 adhesives, however, there are 2 missing values for peel (curves 277


and 279, see Table 5 in Appendix 2). Thus, the data set size is N = 26 for loop tack, and only

N = 24 for peel. For the 16 first adhesives, the loop tack and the peel values {yk} are the mean

of Mk = 3 repeat measurements {yk,j}, but only the means {yk} are available (see Appendix 2,

Table 5). For the last 10 adhesives (shaded cells in Table 5), the repeat measurements {yk,j} are

available, and the loop tack and the peel values are the mean of Mk = 4 or 5 measurements. For

adhesive k, with k=17 to 26, the estimate (sk)2 of the measurement variance (sk)2 is given by:

s k 2 =

y k,j � yk 2!j=1

M k

Mk � 1

The values of the {sk} are shown in Figure 2.

0 5 10 15 200

1

2

3

2841

2871

3011

3041

19381939 1940

1941

1942 1943

mean

0 5 10 15 200

1

2

3

2841

2871

3011

3041

19381939 1940

1941

1942

1943

std

(N

/25 m

m)

std

(N

/25 m

m)

mean

a)

b)

Figure 2. Local estimates {sk} of the standard deviations of the measurements obtained with repeats

for the last 10 adhesives of Table 5, and mean estimate s (dashed line): a) loop tack, b) peel.

The measurement noise is clearly heteroscedastic, i.e. its variance depends on the adhesive. In

principle, this should lead to a weighted least squares approach. However, since the {yk,j} and

hence the {sk} are available for only 10 adhesives out of 26, such an approach is not feasible in

practice, and the model parameters will be estimated using ordinary least squares, see Appendix

3.

In the following, we will use the average estimate of the standard deviation of the output. As a

matter of fact, if the variances are all assumed equal to a common value s2, an estimate s2 of s2

is computed on the available repeat measurements:

s 2 =

(Mk � 1) s k 2!k=17

26

M � 10 with M = Mk!

k=17

26


Numerically, we obtain s = 1.19 N/25 mm for loop tack and s = 1.23 N/25 mm for peel, see

Figure 2. These values will be used later for the selection of the appropriate model complexity

with a lack of fit test, see section III.4 and Appendix 3.

III.2. Choice of the candidate descriptors

As stated in the introduction, it is assumed that the information about the loop tack and the peel

properties is contained in the force-displacement curves measured during a probe test. Since the

displacement is the same for each test and hence for each adhesive, only the force as a function

of time is useful for the prediction of the loop tack and of the peel values.

0 10 20 30 40 50 60

-1

0

1

time (s)

forc

e (

N)

a)

0 20 40 60 80 100

-1

0

1

descriptor number

forc

e (

N)

b)

0 10 20 30 40 50 60

-1

0

1

descriptor number

forc

e (

N)

c)

Figure 3. Thick curve of Figure 1 (N°300): a) force measurements performed every 2 ms; b) 100

candidate descriptors; c) 60 candidate descriptors.

In the following, the dimension of the descriptor space is the number of time instants at which

the force values are considered, between the contact creation and the time corresponding to the

longest test of the dataset. The force measurements are performed every 2 milliseconds, see

Figure 3a which shows one of the curves obtained for the adhesive N°300, the thick curve of

Figure 1. We performed a dimension reduction by drastically reducing the time resolution.


Hundred descriptors, i.e. force values, seem enough to grasp the information contained in the

force curve, see Figure 3b (each descriptor value is the mean of the 330 corresponding force

values). We even consider lower resolutions, see Figure 3c. There is a compromise between

losing information and reducing the number of candidate descriptors: as shown in section IV, a

good compromise is achieved for a number of descriptors equal to 60.

As stated above, the reproducibility of the probe test is quite high. However, the curves of

some adhesives show a significant variability, especially at the end of the curve see Figure 1.

Therefore, we choose to consider the mean of the previous descriptor values over the repeats of

the probe test (Mk = 3 to 5 repeats).

An alternative choice of descriptors would be that of physically significant characteristics of the

curves, such as the force minima and maxima and the corresponding time instants, the area of

the curve below and above zero, etc. (see Appendix 4). We also experimented such a choice,

but obtained worse results than with the raw force descriptors.

III.3. Construction of nested models involving the most relevant descriptors

We are now in the presence of the descriptors judged potentially relevant for the prediction of

the loop tack or of the peel. We denote the candidate descriptors values by the {xi}i=1 to n, the

components of the n-descriptor vector x, and by y the mean value of the loop tack or of the

peel. The assumption that the information about the property is contained in the candidate

descriptors is equivalent to assuming the existence of a model of the form:

y = g(x) + w (1)

where g is an unknown function, and w is a zero mean random variable modeling the output

noise studied. The goal is hence to choose a suitable parameterized function f(x, q) and a value

qopt of the parameters q such that f(x, qopt) is close to g(x) in the domain of interest.

A first issue is that, the dimension of x being large (n � [50; 100]), even a simple affine model

involves too many parameters (n+1) given the number of measurements (N = 26 for tack and

24 for peel). Thus, the design of the model heavily relies on the selection of the relevant and

non redundant descriptors among the candidates. Therefore, the model must be mathematically

suitable for the selection of the most relevant descriptors.

Next, the model should also be suitable for the modeling of interactions between different parts

of the curve and of possible nonlinearities.

Polynomial models satisfy both conditions: they are linear in their parameters, but nonlinear in

the descriptors. A polynomial of degree two is able to represent linear (monomials xi) and

quadratic (monomials (xi)2) dependencies, as well as interactions between different regions of

the curve (cross-products monomials xi xj). The value of the degree is often limited to 2 by the

available computer memory1.

1 As a matter of fact, the polynomial of degree d possesses a number of monomials equal to :


A third issue is how to construct polynomials involving an increasing number of the most

relevant monomials: this is achieved using the modified Gram-Schmidt orthogonalization

procedure [Golub 1996], the question of when to stop this construction being answered in the

next subsection. Let us consider a polynomial of fixed degree d of the n candidate descriptors

{xi}. For example, if the list of the candidate descriptors is x1, x2, �, xn, the polynomial of

degree 2 involves a constant term and the monomials: x1, x2, �, xn, x1 x2, x1 x3, �, xn-1 xn,

x12, �, xn

2. We denote the N-vector of the mean loop tack or peel values by y, and the N-

vectors corresponding to the monomials by the {xi}.

y

x1

x2x3

x4

x5

RN

Figure 4. The vector {x4} is such that |cos(y, xj)| is the largest, and is introduced first.

The monomials {xi} are introduced according to their decreasing contribution to the explanation

of y. The monomial j that is considered to have the maximum contribution to the output is the

monomial such that |cos(y, xj)| is the largest; it is also the monomial which decreases most the

residual sum of squares, see Appendix 3. The remaining monomials and the output are

orthogonalized with respect to this first monomial j using the Gram-Schmidt orthogonalization

algorithm. The procedure is repeated in the subspace orthogonal to the first ranked monomial j,

and so on.

An alternative to this model construction could be a principal component analysis (PCA)

followed by the construction of least squares models having the first principal components as

inputs. Such an approach has two main disadvantages. First, PCA is an unsupervised method:

the principal components are computed from the descriptor values only, without using the

output values of the propery to predict. Thus, the first principal components have no reason to

Nmono(n, d) = Kni!

i=1

d

= Cn+i-1i

!i=1

d

where Kni is the number of i to i combinations of n objects with repetitions, and Cn

i = i!/n!(n-i)! is the number

of i to i combinations of n objects without repeats. For example, the polynomial of degree d = 2 of n = 100

input variables possesses 5 150 monomials, that of degree d = 3 possesses 176 850 monomials.


explain the property, whereas, in our procedure, the monomials are introduced precisely in the

order of their decreasing correlation with the output property. Next, PCA is a linear method: the

principal components are linear combinations of the descriptors. Thus, PCA is not able to

capture the influence of nonlinearities or of interactions between descriptors, whereas higher-

order monomials do.

In the next subsection, we study when to stop the Gram-Schmidt model construction.

III.4. Selection of the model with minimal necessary complexity

The model with minimal necessary complexity is the model with the smallest number of

monomials which achieves a precision, i.e. a mean square error, that is comparable to the

variance l2 of the output. Each time a monomial is introduced, we can test whether the mean

square error of the corresponding polynomial is not too large, using a lack of fit test, whose

principle is detailed in Appendix 3. The first polynomial which satisfies the test is selected. The

test is performed with a type I risk a = 1%. Since there are Mk = 3 to 5 repeats for each

adhesive, we make the approximation that Mk = m = 4, i.e. that l2 = s2/m. This corresponds to

an estimate of l equal to v = s/2, i.e. v = 0.59 for loop tack and v = 0.61 for peel.

In order to quantify the significance of the force descriptors involved in the selected model, we

define the sensitivity of its output with respect to each force descriptor as:

dj = 1N

"f(x,qLS)

"xj k !

k=1

N

= 1N

qLSi "xi

"xj k!i=1

p

!k=1

N

for j=1 to n

where N denotes the number of PSAs, f(x, qLS) is the model output, n is the number of

descriptors, p is the number of selected monomials, xi is the i-th selected monomial, and qLSi is

the corresponding least squares parameter estimate.

III.5. Model performance estimation

The performance index of the model is defined as its mean error over all possible adhesives in

the family of interest, i.e. as the root of the expectation of its mean square error E((y �

f(x,qLS))2).

Since the data set is small, we provide a leave-one-out (LOO) cross validation estimate of this

performance index. The LOO estimate for a model obtained with N adhesives is defined as

follows: the k-th LOO error is the error for the k-th adhesive made by the model obtained when

the k-th adhesive is left out for the model parameter estimation, and the LOO estimate is the root

of the mean of the N squared LOO errors. These LOO errors can be computed conveniently as

functions of the residuals of the model obtained on the whole data set, see Appendix 3. The root

mean square error (RMSE) obtained on the whole data set is denoted by RMSEwhole, and that

of LOO by RMSELOO.

We also provide the RMSE obtained when estimating the parameters of the model on a training

set of 20 adhesives for tack and 18 adhesives for peel, and testing the model on an independent


test set of 6 adhesives corresponding to a series of different formulations (see Appendix 2). The

RMSE obtained on the training set is denoted by RMSEtrain, and that on the test set by

RMSEtest. A good model is such that RMSELOO/RMSEwhole and RMSEtest/RMSEtrain are not

much larger than 1.

IV. Application to the prediction of the loop tack and peel values

IV.1. Using the lack of fit test

For loop tack, with a degree 1, 15 monomials are needed for the lack of fit (LOF) test, but the

RMSELOO and the RMSEtest are large. With a degree 2, only 11 monomials are needed, and all

the RMSEs are acceptable. For peel, with a degree 1, 16 monomials are needed for the LOF

test, but the RMSELOO and the RMSEtest are also large. With a degree 2, only 9 monomials are

needed, and all the RMSEs are acceptable. The performance of the tack and peel models of

degree 2 is summarized in Table 1. We note that the models obtained with the LOF test possess

either a RMSELOO or a RMSEtest that is large compared to the RMSEwhole or to the RMSEtrain.

outputproperty

nbmonomials/descriptors

nb PSAstraining/test

output s.e.estimate v(N/25 mm)

RMSEwhole

(N/25 mm)RMSELOO

(N/25 mm)RMSEtrain

(N/25 mm)RMSEtest

(N/25 mm)

loop tack 11/16 20/6 0.59 0.39 1.44 0.40 0.54peel 9/14 18/6 0.61 0.64 1.17 0.49 1.44

Table 1. Model of degree 2 selected with the LOF test (100 descriptors).

The model sensitivities are shown in Figure 5. The 11 monomials of the tack model are, in the

order of relevance: x2 x8, x6 x100, x6 x70, x40 x85, x2 x31, x59 x65, x39 x84, x41 x70, x6 x8, x44

x57, x252. The 9 monomials of the peel model are: x59

2, x5 x8, x6 x75, x59, x23 x74, x75 x78, x9

x24, x44 x45, x30 x57. Most of the monomials are hence cross-product monomials. This

illustrates the benefits of the statistical approach, whose only assumption is that the information

is contained in the raw descriptors: it discovers interaction effects between different parts of the

curve which were not predictable a priori, and which are still difficult to analyze.

Note also that the tack model involves descriptors located at the end of the curves, i.e.

descriptors whose variability is generally high, whereas the peel model does not. Though it may

seem surprising, the fact that relevant descriptors are located after the contact of certain tests is

over is not pathological. It simply means that these descriptors will not contribute to the output

of the model for these tests, whereas for other tests, they will contribute (positively or

negatively).

The number of parameters of the models selected with the LOF test are too large given the

training set size (N=26 or 24), and the discrepancy between the RMSEwhole and the RMSELOO

are too large (a factor of 4 for tack, of 2 for peel). As a matter of fact, when increasing the


number of descriptors, the correlation will always be better on the training set, but not

necessarily on unseen PSAs. Furthermore, we do not have total confidence in the output

variance estimate used for the LOF test, since the detail of the measurements is unknown for 16

adhesives out of 26.

0 20 40 60 80 1000

20

40

descriptor number

tack s

ensitiv

ity

0 20 40 60 80 1000

1

2

3

descriptor number

peel sensitiv

ity

Figure 5. Sensitivities of the models selected with the LOF test (100 descriptors).

IV.2. Achieving a bias/variance compromise

Thus, we retained models with less monomials than those selected with the LOF test, but for

which the discrepancy between the RMSEwhole and the RMSELOO is smaller. In other words,

we pay for reducing the model variability (variance) by accepting a lack of fit (bias). For tack,

we retain a model with 9 monomials, and for peel, a model with 7 monomials, see Table 2. The

discrepancy between the RMSEwhole and the RMSELOO is smaller, and the number of

parameters is more reasonable. However, the RMSELOO and the RMSEtest of the tack model are

still large as compared to the output standard error estimate v.

outputproperty




RMSEwhole

(N/25 mm)RMSELOO

(N/25 mm)RMSEtrain

(N/25 mm)RMSEtest

(N/25 mm)

loop tack 9/13 20/6 0.59 1.26 2.86 1.07 2.43peel 7/10 18/6 0.61 0.93 1.24 0.83 1.20

Table 2. Model of degree 2 achieving the best bias/variance compromise (100 descriptors).

The sensitivities are shown in Figure 6. The 9 monomials of the tack model involve 13

descriptors, and the 7 monomials of the peel model involve 10 descriptors. The tack model

again needs descriptors located at the end of the curves.


0 20 40 60 80 1000

10

20

30

descriptor number

tack s

ensitiv

ity

0 20 40 60 80 1000

1

2

descriptor number

peel sensitiv

ity

Figure 6. Sensitivities of the models achieving a bias/variance compromise (100 descriptors).

IV.3. Achieving a bias/variance compromise and using only 60 descriptors

The results are quite stable when the number of descriptors is in the interval [50 ; 100]. We

present the results obtained with 60 descriptors, which are summarized in Table 3.

outputproperty




RMSEwhole

(N/25 mm)RMSELOO

(N/25 mm)RMSEtrain

(N/25 mm)RMSEtest

(N/25 mm)

loop tack 9/14 20/6 0.59 1.38 2.19 1.30 2.00peel 7/11 18/6 0.61 0.88 1.14 0.79 1.28

Table 3. Model of degree 2 achieving the best bias/variance compromise (60 descriptors).

The 9 monomials of the tack model are, in the order of relevance: x2 x54, x4 x39, x3 x4, x27,

x33 x49, x1 x58, x4 x33, x22 x46, x16 x35. The 7 monomials of the peel model are: x41 x43, x3 x5,

x38 x43, x3 x34, x452, x28 x48, x35 x42. Most of them are still cross-product monomials.

0 10 20 30 40 50 600

10

20

tack s

ensitiv

ity

0 10 20 30 40 50 600

20

40

descriptor number

peel sensitiv

ity

Figure 7. Sensitivities of the models achieving a bias/variance compromise (60 descriptors).


2841 1938

285

277 1941 1942

302

1940 1939

0 5 10 15 20 25

0

5

10

15

20

25

243

244

279

282

283

284

286

287

299

300

301

303

304

2871

3011

3041

1943

measured tack

pre

dic

ted t

ack

Figure 8. Predicted/measured tack, for 20 training (crosses) and 6 test adhesives (filled circles).

Figure 7 displays the model sensitivities. Figures 8 and 9 display the outputs of the models for

the training and test adhesives. The LOO and test errors of the tack model are even smaller than

with 100 descriptors, probably due to the averaging effect on the end of the curve.

303

285

3011

302 1940

1941 1942

1938 1939

0 5 10 15 20

0

2

4

6

8

10

12

14

16

18

20

243

244

282

283

284

286

287

299

300

301

304

2841

2871 3041

1943

measured peel

pre

dic

ted p

eel

Figure 9. Predicted/measured peel, for 18 training (crosses) and 6 test adhesives (filled circles).


V. Discussion

Let us first summarize the results. Table 3 shows that the peel model achieves RMSELOO and

RMSE test twice as large as the standard deviation, with a reasonably small number of

parameters as compared to the available number of adhesives. The tack model is less accurate,

with RMSELOO and RMSEtest three times as large as the standard deviation, despite the larger

number of parameters. The loop tack may be more difficult to predict for two reasons. First

because the relationship between the descriptors and loop tack may be more nonlinear. Second

because the tack model involves descriptors located at the end of the curves, where the probe

test is less reproducible, whereas the peel model does not.

Aside from this assessment of the performance of the model from a purely statistical point of

view, we can discuss in more details the advantages and limitations of the approach as

presented here. A number of points need to be addressed:

- The adhesives for both the training set and the test set were all of the same family of

adhesives (SIS + tackifying resins). It remains to be seen whether a model developed for a

specific family of PSA would work to predict peel and loop test for a different family of

PSA. Our guess is that as long as the interactions between the PSA and surface do not vary

much, the model should be applicable to another family of adhesives such as acrylics.

However more tests would be needed to confirm that.

- We did not discuss in this paper the influence of debonding mechanisms on the peel force

values. It is well known that in a certain range of peel velocities, many PSA experience a so-

called stick-slip debonding behaviour. In this case the peel force is no longer defined as a

single value, but rather two values (an initiation force and an arrest force, which depend

among other things on the details of the experimental setup). Clearly this type of mechanism

cannot be predicted by our model, but one could envision that a different set of descriptors

could be used to build a model able to predict the occurrence of stick slip.

- Although indeed the model does not use physically meaningful descriptors, it certainly

assumes that the information necessary to predict the peel force or loop test is contained in

the probe test curve. It uses in other terms the probe test curve as a signature of the material

(like an infrared spectrum) and extracts the relevant bits of the signature to predict the peel

force or loop test result. For this approach to be successful, it is important to minimize the

variability of all the parameters that are not the properties of the adhesive itself. This means

that all adhesive layers should have the same thickness and surface roughness, that the probe

used for the probe tests should be the same for all tests, that the experimental conditions for

the probe test (probe velocity, contact force and contact time) and for the loop test and peel

test should be kept constant.


VI. Conclusion

The results we obtained confirm our starting hypothesis that the information about the loop tack

and the peel is indeed contained in the force signal obtained with a probe test. Although the

accuracy of the models is not yet optimal, we feel that the proposed procedure is very promising

and is likely to provide increasingly more refined predictive models of standard PSA properties.

This approach is complementary to an approach based on intuitively identifying the most

relevant physically significant descriptors of the force-displacement curves such as the

maximum stress or the adhesion energy. It is worthwhile to note that the most relevant

descriptors found by the empirical model are often coupled and may provide ideas for more

complex and better physically based descriptors.

As usual with empirical models, the quality of the model depends on the range and

representativity of the data available to construct it, and we expect further improvements from a

reliable estimation of the measurement variability for each adhesive, which will allow a correct

weighting of the data points, and from the use of a larger, more representative data set. Such a

model would have a great potential as a development and screening tool. Probe tests results,

obtained easily and rapidly even from rather small amounts of materials, would then guide the

synthetic chemist to select the most promising synthesis conditions, and to perform a time-

consuming complete set of tests on these selected formulations only.

We also feel that, because it probes the large strain nonlinear elastic properties of the materials,

the probe test measurements may contain some of the information necessary to predict the

resistance to shear. This is an industrially very important but notoriously difficult property to

predict from the molecular structure of the adhesive, and we plan to apply the same statistical

methodology in that direction.

VII. Acknowledgements

We are very grateful to Hilde De Muyck for producing the results that motivated this study, as

well as to Jan Vaneecke, who performed the major part of the experimental work.


Appendices

A1. Adhesive formulations

PSA N° Formulation

243 D1161/Piccotac 212 : 12 / 5

244 D1161/Piccotac 212 : 12 / 10

277 D1161/Piccotac 212 : 12 / 8

279 D1161/Piccotac 212 : 12 / 12

282 D1161/Piccotac 95 : 12 / 5

283 D1161/Piccotac 95 : 12 / 10

284 D1161/Piccotac 95 : 12 / 20

285 D1161/Foral 85 : 12 / 5

286 D1161/Foral 85 : 12 / 10

287 D1161/Foral 85 : 12 / 20

299 D1160/Piccotac 95 : 12 / 5

300 D1160/Piccotac 95 : 12 / 10

301 D1160/Piccotac 95 : 12 / 20

302 D1160/Foral 85 : 12 / 5

303 D1160/Foral 85 : 12 / 10

304 D1160/Foral 85 : 12 / 20

2841 D1161/Piccotac 95E : 12 / 20

2871 D1161/Foral 85: 12 / 20

3011 D1160/Piccotac 95E : 12 / 20

3041 D1160/Foral 85: 12 / 20

1938 D1161/Foral 85 : 12 / 8

1939 D1161/Foral 85 : 12 / 12

1940 D1160/Foral 85 : 12 / 8

1941 D1160/Foral 85 : 12 / 12

1942 D1160/Piccotac 95 : 12 / 8

1943 D1160/Piccotac 95 : 12 / 12Table 4. Formulations of the 26 adhesives of the data set.


A2. Description of the data set

PSA

N°

number of

curves

Mk (loop

tack)

loop tack (N/25

mm)

mean yk (std sk *)

Mk (peel) peel (N/25 mm)

mean yk (std sk *)

243 5 3 8.8 (-) 3 8.9 (-)

244 4 3 18.1 (-) 3 17.1 (-)

277 3 3 8.6 (-) 0 �

279 4 3 2.4 (-) 0 �

282 5 3 12.8 (-) 3 10.2 (-)

283 3 3 17.8 (-) 3 16.4 (-)

284 4 3 0.8 (-) 3 13.6 (-)

285 5 3 8.5 (-) 3 7.2 (-)

286 5 3 15.5 (-) 3 9.1 (-)

287 4 3 25 (-) 3 14.6 (-)

299 3 3 12 (-) 3 8.4 (-)

300 4 3 21.2 (-) 3 15.5 (-)

301 4 3 0.2 (-) 3 1.8 (-)

302 3 3 5.8 (-) 3 5.5 (-)

303 3 3 12.4 (-) 3 8.4 (-)

304 4 3 26.6 (-) 3 13.4 (-)

2841 3 5 1.8 (0.6) 4 19 (1.3)

2871 3 5 15.4 (1.2) 5 11.3 (0.5)

3011 3 5 0.3 (<0.1) 5 17.4 (3.3)

3041 3 5 14.5 (2.5) 5 11.2 (1.1)

1938 5 4 3.9 (<0.1) 5 6.0 (0.4)

1939 5 4 8.3 (0.5) 5 7.9 (0.4)

1940 5 5 4.1 (0.3) 5 6.9 (0.3)

1941 5 4 9.4 (0.4) 5 8.8 (0.3)

1942 5 5 9.5 (0.9) 5 10.5 (0.4)

1943 5 4 10.6 (2.1) 5 13.0 (0.4)

s = 1.19 s = 1.23Table 5. The available data set. (*) Standard deviations estimated with repeats, if available. The shaded

cells correspond to the adhesives for which the standard deviations are available.


A3. Least squares and statistics

We consider a data set of N adhesives represented by the couples {xk, yk}k=1 to N. The {xk} are

the force values at n sampling instants, and the {yk} are the corresponding mean values of the

considered property, loop tack or peel. To assume that the information about the property is

contained in the candidate descriptors {x} is equivalent to assume the existence of the model:

y = E(y) + w = g(x) + w

where the mean g is an unknown function of x, and w is a zero mean random variable modeling

uncertainty. We further assume homoscedasticity, i.e. that the outputs {yk} are uncorrelated

and have the same variance l2.

Polynomial model

The goal is to choose a suitable parameterized function f(x, q) and a value qopt of the parameters

q such that f(x, qopt) is close to g(x) in the domain of interest. Here, we consider a polynomial

model with p monomials and parameter p-vector q. We denote the monomials of adhesive k by

xk, which is a p-vector, and whose first component is equal to one. For example:

f(x k, q) = q1 + q2 x15k + q3 x45

k 2 + q4 x10

k x31k = q1 x1

k + q2 x2k + q3 x3

k + q4 x4k = x k T q

The estimates of this model for the N adhesives of the training set are hence given by:

x 1 T

�

x N T

q # 1 x2 � xp q # X q

where the {xi} are N-vectors. The (N ,p) matrix X is known as the experiment matrix.

Assuming that N $ p and rank(X) = p, i.e. that the {xi} are linearly independent, the least

squares (LS) parameter estimate qLS minimizing || y � X q ||2 is given by:

qLS = X T X-1

XT y

The residual N-vector r is defined as the N-error vector of the LS model, that is:

r = y � X qLS

The residual sum of squares is equal to rT r.

Null hypothesis

Let us consider the hypothesis that there exists an unknown parameter vector q* such that:

H0 : y = E(y) + w = X q* + w

and where hence E(w) = 0. This hypothesis is called the null hypothesis (H0): it is the

hypothesis that the monomials {xi} are sufficient to explain the output property y, i.e. that the

model is unbiased. In other words, the subspace spanned by the {xi} (the image of the matrix

X) contains the mean of y. If the monomials {xi} are not sufficient, H0 is false, which means

that the model is biased. Since we search for an unbiased model, H0 is the hypothesis we want

to test.


If H0 is true, E(qLS) = q*. Moreover, the noise being homoscedastic, the residuals provide an

unbiased estimate of the variance l2 of the outputs {yk}:

E r T rN � p

= l2

Lack of fit test

Suppose that we dispose of an independent estimate v2 of the variance l2 that is reliable, i.e.

unbiased: E(v2) = l2. Intuitively, if H0 is true, the residual vector r is essentially due to

uncertainty and the following ratio f should be close to 1:

f =

r T rN � p

v 2

On the other hand, if H0 is false, the model is too simple (biased) and r is not only due to

uncertainty but also to a model bias, and the ratio f will be much larger than 1.

If we assume Gaussian outputs, it is easy to test whether f is significantly larger than 1, see

[Draper & Smith 1998] [Rivals & Personnaz 2003b]. Then, if v2 is obtained from q

independent values:q v2

l2 % c2(q)

where c2(q) denotes the Pearson distribution with q degrees of freedom. If moreover H0 is

true, we have:r T r

l2 % c2(N � p)

and hence:

f =

r T rN � p

v 2 % Fisher(N�p, q)

This leads to the lack of fit (LOF) test. Let us choose a value a for the probability of rejecting

H0 when it is true, the type I risk. Then, we will decide to reject H0 when:

f =

r T rN � p

v 2 > FN�p, q(1�a)

where FN�p,q(1�a) denotes the 1�a quantile of the Fisher distibution with N�p and q degrees of

freedom. In other words, if the value of f is too large given a, we will decide that the model is

biased with a risk a of being wrong.

Independent variance estimate

The available individual measurements {yk,j}k=17 to 26 of the property are assumed to have a

common variance s2. Hence, we have the following unbiased estimate s2 of s2:

s 2 =

(Mk � 1) s k 2!k=17

26

M � 10 =

y k,j � yk 2!j=1

M k

!k=17

26

M � 10


where, for each adhesive k, Mk is the number of repeats, yk,j is the value of the j-th repeat, yk

is the mean value over the repeats, and M = Mk!k=17

26.

Assuming Gaussian measurements:

s 2 M�10

s 2 % c2(M�N)

The outputs being a mean of Mk = 3 to 5 repeats, we choose to approximate Mk � m = 4.

Thus, we can consider l2 = s2/m as the common variance of the outputs {yk}, and v2 = s2/m

as its estimate, obtained with M� 10 degrees of freedom.

Leave-one-out errors

The LS parameter vector qLS of the model is obtained using the whole data set of size N. The

residual vector of this model is r = [r1 � rN]T = y � X qLS . The k-th leave-one-out (LOO)

error is the error made on adhesive k by the model obtained when adhesive k is left out from the

data set. The latter does not need to be computed, since the k-th LOO error is a simple function

of the k-th residual rk:

eLOOk = rk

1 � hk

where hk is the k-th diagonal term of the (N,N) orthogonal projection matrix X (XT X)-1 XT on

the image of X, or �hat� matrix, see [Rivals & Personnaz 2003a].


A2. Alternative choice of physical descriptors

We constructed 18 descriptors of the shape of the force signal as a function of time. These new

inputs are the following :

1) The area below zero.

2) The area above zero.

3) The initial slope of the curve.

4) The slope of a line joining the point at which the force reaches �1N and the point at which the

force reaches its first maximum.

5) The slope of a line joining the point at which the force reaches its first maximum and the

point at which the force reaches the next minimum.

6) The slope of a line joining the point at which the force reaches this minimum and the point at

which the force reaches the second maximum.

7) The slope of a line joining the point at which the force reaches its second maximum and the

point at which the force reaches zero.

8) The instant when the force reaches �1N.

9) The instant when the force crosses zero.

10) The instant when the force reaches its first maximum.

11) The instant when the force reaches the next minimum.

12) The instant when the force reaches its second maximum.

13) The instant when the force reaches zero.

14) The value of the first maximum of the force.

15) The value of the next minimum of the force.

16) The value of the second maximum of the force.

17) The number of force maxima.

18) The curvature at the minimum.

10 20 30 40 50 60

-1

0

1

2

time (s)

forc

e (N

)

Figure 10. Some of the descriptors.


References

K. Brown, J. C. Hooker, and C. Creton, Macromolecular Materials and Engineering 287,

163-179 (2002).

H. K. Chuang, C. Chiu, and R. Paniagua, Adhesives Age, 18-23 (1997).

C. Creton, J. C. Hooker, and K. R. Shull, Langmuir 17, 4948-4954 (2001).

C. Creton, MRS Bulletin 28, 434-439 (2003).

A. J. Crosby and K. R. Shull, Journal of Polymer Science: Part B: Polymer Physics 37, 3455-

3472 (1999).

N. R. Draper and H. Smith. Applied regression analysis. John Wiley & Sons inc., New York

(1998).

G. H. Golub and C.F. Van Loan. Matrix computations. John Hopkins University Press,

Baltimore, third edition (1983).

J. C. Grunlan, D. L. Holguin, H. K. Chuang, I. Perez, A. Chavira, R. Quilatan, J. Akhave,

and A. R. Mehrabi, Macromolecular Rapid Communications 25, 286-291 (2004).

V. Kulikov, V. M. Mirsky, T. L. Delaney, D. Donoval, A. W. Koch and O. S. Wolfbeis,

Measurement Science and Technology 16, 95-99 (2005).

H. Lakrout, P. Sergot, and C. Creton, Journal of Adhesion 69, 307-359 (1999).

H. Lakrout, C. Creton, D. Ahn, and K. R. Shull, Macromolecules 34, 7448-7458 (2001).

PSTC, Test methods for pressure sensitive adhesive tapes, 13th ed. (Pressure Sensitive Tape

Council, 2000).

I. Rivals and L. Personnaz, Journal of Machine Learning Research 3, 1383-1398 (2003a).

I. Rivals and L. Personnaz, IEEE Transactions on Neural Networks 14, 804-819 (2003b).

A. Roos and C. Creton, Macromolecular Symposia 214, 147-156 (2004).

K. R. Shull and C. Creton, Journal of Polymer Science: Part B: Polymer Physics 42, 4023-

4043 (2004).

H. Zhang, R. Hoogenboom, M. A. R. Meier and U. S. Schubert, Measurement Science and

Technology 16, 203-211 (2005).

A. Zosel, Colloid and Polymer Science 263, 541-553 (1985).

a statistical method for the prediction of the loop tack

Documents